基于量化的CNN恶意代码检测方法

CNN Malicious Code Detection Method Based on Quantization

针对当前采用深度学习的恶意代码检测方法存在计算资源消耗较大,难以部署到资源受限的边缘设备的问题,提出了一个对恶意代码可视化检测模型进行压缩的方法。将恶意软件转换成灰度图像后,输入卷积神经网络,对卷积神经网络的输入值和权重进行量化运算,构建恶意代码的分类模型HBF-VGG14-Net,通过量化后的VGG14模型进行训练及测试得到检测结果。实验结果表明,相比全精度VGG14模型,HBF-VGG14-Net在精度损失较小的情况下,能够实现模型压缩28倍。所提方法预处理操作简单,训练的模型占用内存较小,故HBF-VGG14-Net模型可用于边缘设备的恶意软件检测。

Aiming at the problem that the current malicious code detection method based on deep learning consumes large computing resources and is difficult to deploy to resource-constrained edge devices,a method to compress the visual detection model of malicious code is proposed.Firstly,the malicious software was converted into a grayscale image,which was then input into a convolutional neural network.The input values and weights of the convolutional neural network were quantified,and a malicious code classification model HBF-VGG14-Net was constructed.The detection results are obtained by training and testing the quantized VGG14 model.Compared with the fullprecision VGG14 model,HBF-VGG14-Net can achieve the model compression of 28 times with less precision loss.The preprocessing operation of this method is simple,and the trained model occupies less memory,so the HBFVGG14-Net model can be used for malware detection of edge devices.